Emotion dysregulation is a broad, transdiagnostic risk and maintenance factor for numerous psychological disorders. Current treatments targeting emotional regulation are only partly effective, likely due to suboptimal delivery systems. In this study, we take the smartphone-based just-in-time, adaptive intervention approach, allowing real-time data collection and in-the-moment prediction of mood or behavioral states of interest. Various sensor data such as heart rate, physical activity, skin temperature and electrodermal activity are collected. The combination of multiple wearable sensor data provides a more comprehensive picture of the research problem at hand. However, it also requires the development of new statistical methods that can make full usage of the informational complexity. A large number of physiological features are extracted from different wearable sensors in both time and frequency domains. We then develop a new data integration strategy based on multiple kernel learning to combine features from different sensors to predict affect lability. Analysis results show that the proposed approach can effectively and efficiently detect urges and emotional eating episodes.